Corpus Selection

For this research project, I selected 7 Dutch and 7 American hip hop artists. Besides comparing the two genres, this enables us to look at the characteristics of each individual artist. For each artist Spotify has created a ‘This is …’ playlist with around 50 songs. I noticed a lot of collaboration between artists. This might result in songs appearing in multiple playlists, which might have a misleading effect. To be able to do a good comparison, I decided to select only the songs by the artist himself. I hope this gives a good representation of the style and properties of the artists. After making this selection, the playlists differed a lot in size, but I accepted this because the number of songs for the dutch and American genre were almost the same. This analysis is focussed on relatively new hip hop, with songs uploaded from 2017 till now.

Spotify Playlists: Nederhop

Spotify Playlists: American hip hop

First look at the data

Findings

These are the means and standard deviations for the numeric features. The featue values for American and Dutch hip hop are alike. It can be noticed that the standard deviations for the Dutch hip hop are smaller. This was surprising, because in the literature, Nederhop seemed to be a broader genre.

Russell’s circumplex model

Findings

On the left we see the data points in an enery valence space. This space can be linked to different emotions. Above, Russell’s circumplex model is displayed. In this model, the vertical axis represents arousal, which for now we consider comparable with energy. The horizontal axis represents valence. The middle represents medium valence and medium arousal.

Most data appears in the top half. This makes sense, considering hip hop is more happy or angry than relaxed or sad. American hip hop appears in the sad section more often. When looking at artists individually, we don’t see any clear clusters.

Findings

To define the danceability of a song, Spotify considers tempo, rhythm stability, beat strength, and overall regularity. As hip hop is a huge dance culture, it is expected that the songs are often very danceable. This is visible in the data. I didn’t think acoustic hip hop was very plausible, so I expected to find some outliers. I don’t think we should consider “Grand Piano” by Nicki Minaj hip hop, but I do think “Lust For Life” by Drake is a great example that acoustic hip hop is possible.

Findings

In the previous page we looked at the danceability of songs. Although danceability is not a very straight forward feature, I want to look for other things that might indicate the dancebility of songs. I selected “Gemaakt voor dit” by Lil’ Kleine and “Signs” by Snoop Dogg, because they are extremely danceable according to Spotify. After listening to the songs, I must say I agree.

For each song, the left self-similarity matrix is based on timbre, and the right one is based on chroma. The features were summarised at bar level and the axes are in seconds.

In the matrices, we see a checkerboard structure. These blocks represent homogeneous regions in the songs. The diagonal lines, parallel to the main diagnal represent exact repetions in the music. Both songs show a lot of repitions, which makes sense. To make it easy to dance to a song, I think the song should be predictable and that can be accomplished by repetition. The chroma and timbre patterns for “Gemaakt voor dit” seem to fit nicely on top of each other, showing that homogeneous regions in chroma correspond to homogenious regions in timbre.

The keygram for “Signs” does not seem very plausible, because it shows high intensity for almost all keys at the same time.

Findings

I randomly selected 30 Dutch songs and 30 American songs to perform unsupervised learing on. On the left, we see an hierarchical clustering for the data.

The heatmap shows the intensity in which the features contributed to the clustering. As the clusters are all in the blue region, it is hard to point out features that are most responsible. Instrumentalness was apparantly very important for the song ‘Heard Em Say’, assiging it to a cluster of its own.

Findings

This section shows two classifation tasks: genre classification and artist classification. In both cases a random forest classifier was used, generating 500 trees. To reduce training time, for each genre 75 songs were sampled and 15 songs were sampled per artist. The confusion matrices were obtained by 5-fold cross-validation.

Though there were not very clear clusters visible in the data separating Dutch and American music, a classifier was build to see it could do a better job. This resulted in an accuracy of 0.7933333, a precision of 0.7972973 and a recall of 0.7866667. The next graph shows the feature importance for the random forest classifier. Two of the most important features are plotted. These can differ every time we run the algorithm, because the decision trees are random.

For the artists individually I did not see any clusters at all. A random forest classifier does not perform very well either, as we can see in the confusion matrix for artist classification. Boef, Bizzey and Snoop Dogg are best separable from the rest of the artists.

Conclusions

• Features for American hip hop and Nederhop are alike. In general, American hip hop is a little more spread out.
• Looking at the energy and valence space, hip hop music appears mostly in the angry and happy quadrants.
• Hip hop songs contain a lot of repetitions. This could be one of the reasons for their high danceability.
• Clustering did not deliver any insights about feature importance.
• Features that were most important for classification were duration_ms, timbre component 9 and timbre component 5.
• The accuracy for classifying Dutch and American hip hop was 0.7933333.
• There weren’t a lot of visible feature differences for different artists. This was supported by a poor performance on the classification task for artists.

Discussion

• The data set might contain non hip hop outliers, because during data collection, I focussed on artist rather than genre.
• Separating artists was hard. As I mentioned in the corpus selection, there was a lot of collaboration. It turned out I didn’t filter out all collaboration, because sometimes it is mentioned in the song title (with feat …). Maybe bad performance was caused by a small training set. However, it could also be that artists just are not that different in terms of Spotify features. Removing collaborative data could also have resulted in removing very descriptive songs for distinguishing genre.
• Future research could focuss on outlier detection or try to run a similar analysis on another corpus.

References

Blanchard, B. (1999). The social significance of rap & hip-hop culture. Journal of Poverty & Prejudice, Spring.

Koreman, R. (2014). Legitimating Local Music: Volksmuziek, Hip-Hop/Rap and Dance Music in Dutch Elite Newspapers. Cultural Sociology, 8(4), 501-519.

Spotify. 2019. Spotify Web API. https://developer.spotify.com/documentation/web-api/

Emotion studies in HCI-a new approach - Scientific Figure on ResearchGate. Available from: https://www.researchgate.net/figure/Russells-circumplex-model-of-affect_fig1_229021134 [accessed 19 Mar, 2019] (Image of Russell’s circumplex model)